Google searches accurately forecast RSV hospitalizations

Background: Hospitalization of children with respiratory syncytial virus (RSV) is common and costly. Traditional sources of hospitalization data, useful for public health decision-makers and physicians to make decisions, are themselves costly to acquire and are subject to delays from gathering to publication. Here we use Google searches for RSV as a proxy for RSV hospitalizations. Methods: Searches for "RSV" and numbers of RSV hospitalizations in WA, MD, FL, and CT were examined from 2004--2018. Running correlation coefficients and phase angles between search and hospitalizations were calculated. Various machine learning models were compared to assess the ability of searches to forecast hospitalizations. Using search data from all 50 US states, we use K-means clustering to identify RSV transmission clusters. We calculate the timing of the optimal timing of RSV prophylaxis initiation as the week beginning the 24-week period covering 95% of all RSV cases. Results: High correlations (>0.95) and low phase differences were seen between counts of hospitalizations and search volume in WA, MD, FL, and CT. Searching for RSV began in FL and radiated outward and three distinct transmission clusters were identified: the south and northeast, the northwest and Appalachia, and the center of the country. Calculated initiation dates for prophylaxis closely followed those calculated using traditional data sources (correlation = 0.84). Conclusions: This work validates searches as a proxy for RSV hospitalizations. Search query surveillance of RSV is a rapid and no-cost addition to traditional RSV hospitalization surveillance and may be useful for medical and public health decision-making.